基于高分辨率孪生网络的单目标追踪算法

Single object tracking algorithm based on high-resolution siamese network

针对目前基于孪生网络的目标追踪算法对目标定位不精确,追踪漂移,鲁棒性不强等问题,提出一种基于高分辨率孪生网络的单目标追踪算法.该算法采用高分辨率网络(HRNet)作为特征提取网络,在充分提取图片多尺度特征的同时保留图片的细节信息,然后在孪生网络部分引入"注意力机制",使用SE模块对图片特征进一步精化,突出有用特征.该算法在检测部分融入类似特征金字塔网络(Feature Pyramid Network, FPN)的思想,充分使用深层孪生网络的多尺度特征.在ILSVRC-2015和YouTuBeBB数据集上进行训练时,为了解决训练过程中正负样本和难易样本数量不平衡的问题,本算法使用focal loss作为损失函数,将模型注意力放在更难分辨的正样本上.实验结果表明:该算法在VOT2018测试集上EAO为42.4%.

Aiming at the problems of object tracking algorithm based on siamese network,such as inaccurate object location,tracking drift and weak robustness,single object tracking algorithm based on high-resolution siamese network is proposed.In this algorithm,High-Resolution Network(HRNet)is used as feature extraction network,which can fully extract multi-scale features of the image while retaining the details of the image.Then,attention mechanism is introduced into the siamese network part,and SE(Squeeze-and-Exception)module is used to further refine the image features and highlight the useful features.In the detection part,the idea similar to FPNis incorporated,and the multi-scale features of deep siamese network are fully used.In order to solve the problem of imbalance between the number of positive and negative samples in the training process,focal loss is used as the loss function to focus on the more difficult to distinguish positive samples.The experimental results show that the EAO of this algorithm is 42.4%in VOT2018 test set.